BoilingBench-CV Dataset

Version: v0.1.0 Maintainer: NED3 Laboratory, University of Arkansas License: CC BY 4.0 DOI: 10.5281/zenodo.22264378

Mirror of the Zenodo deposit of 3 September 2026, published here because most users of these data work in the Hugging Face ecosystem. The file set was verified identical to the deposit at upload time: 7,147 files, 4.20 GB uncompressed.

Authors

Hari Pandey (University of Arkansas), Manohar Bongarala (Purdue University), Christy Dunlap (University of Arkansas), Lige Zhang (Drexel University), Justin A. Weibel (Purdue University), Ying Sun (University of North Carolina at Charlotte), and Han Hu (University of Arkansas).

Status and limitations

Read this before reporting results.

• Files inside the package still carry the pre-release version string 0.1.0-internal, and metadata/data_rights_manifest.csv still records release_status = internal package. Both are stale: the package was published as v0.1.0 under CC BY 4.0 with a DOI on 3 September 2026. The files are mirrored here unmodified rather than corrected, so that this copy stays byte-identical to the citable deposit.
• The annotation quality-assurance record (metadata/annotation_qa.md) lists open follow-up items. 156 contours are flagged: 106 with at least one coordinate outside image bounds and 50 with invalid coordinate counts. Flagged contours are marked ignore=1 and excluded from scoring under the v0.1 screening policy; no manual adjudication has been performed.
• The sha256 column of derived/poolboiling-v0.1/manifest.csv is not populated. Per-file checksums for the canonical annotations and the split files are recorded in derived/poolboiling-v0.1/validation.json.
• Some manifest and result files retain absolute paths from the acquisition host. They are provenance records only and are not needed to read the data.

Purpose

Bubble nucleation, growth, coalescence, and departure govern the heat flux a boiling surface can sustain, and measuring those events from high-speed video is the bottleneck in relating surface design to thermal performance. BoilingBench-CV evaluates computer-vision methods for two-dimensional bubble instance segmentation and per-frame bubble morphometry, with explicit attention to whether a method trained in one imaging domain transfers to another.

This package does not establish a tracking or bubble-dynamics benchmark. Per-image contours do not encode reviewed temporal bubble identities, coalescence or breakup events, or departure times.

Contents

Path	Contents
raw/	Source annotated-image subsets in their original annotation formats
derived/	Canonical COCO-style annotations, manifests, splits, screening results
splits/	Immutable grouped benchmark partitions
metadata/	Source, rights, annotation, and quality-assurance records
docs/	Data card and release notes

Canonical pool-boiling benchmark

derived/poolboiling-v0.1/ covers 357 annotated images carrying 16,106 contour records across four regimes:

Regime	Surface	Fluid	Images
FCu-H2O	Copper foam	Water	91
PCu-H2O	Plain copper	Water	91
PSi-HFE	Plain silicon	HFE-7100	119
SSi-HFE	Structured silicon	HFE-7100	56

Raw subsets

Subset	Images	Annotations
BubbleID_base	3,449 JPG	2,957 Labelme JSON
BubbleID_fine-tuned_for_new_facility	24 JPG	24 Labelme JSON
BubbleID-Flow_fine-tuned_for_flow_boiling	130 BMP	130 Labelme JSON
PoolBoilingDatasets	357 JPG	4 paired contour JSON

Model checkpoints are excluded from all subsets; checkpoint licenses may differ from image and annotation rights.

Splits and the leakage rule

The independent unit is a source acquisition group, represented by (regime, source_video). Every frame and its derivatives stay in one partition. Four protocols are provided, each spanning 31 groups: pooled_grouped, water_to_hfe, hfe_to_water, and leave_<regime>. validation.json records an empty leaked_groups list for every split.

Note that source_video is a grouping key. The benchmark operates on still images.

Loading

The repository is a file tree rather than a datasets-loadable config, since annotations follow the COCO convention rather than a Hugging Face schema.

The Dataset Viewer shows derived/poolboiling-v0.1/manifest.csv, the 357-row index of the canonical pool-boiling benchmark. It is a browsable index, not the benchmark data: images live under raw/, and the annotations live in derived/poolboiling-v0.1/annotations.json.

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="UARK-NED3/BoilingBench-CV",
    repo_type="dataset",
)

To pull only the canonical benchmark and skip the large source videos:

path = snapshot_download(
    repo_id="UARK-NED3/BoilingBench-CV",
    repo_type="dataset",
    allow_patterns=["derived/*", "splits/*", "metadata/*", "docs/*",
                    "raw/PoolBoilingDatasets/**/annotatedBubbles/*"],
)

Then read derived/poolboiling-v0.1/annotations.json with any COCO tool, for example pycocotools.coco.COCO.

Domain shift, and how not to over-read it

The water and HFE-7100 regimes differ in more than working fluid. Surface structure, image geometry, optical arrangement, resolution, and facility may all differ between them. A water-versus-HFE performance gap is therefore a compound domain effect and must not be interpreted as a fluid-only causal effect. Report performance by regime, not only as a pooled score.

Reference screening result

A frozen-checkpoint screening on the water_to_hfe test partition (175 HFE-7100 frames) is recorded in derived/poolboiling-v0.1/results/MODEL_COMPARISON.md. No benchmark-label training, threshold tuning, or model adaptation was used.

Model	Pixel IoU	Pixel F1	Instance precision @ IoU 0.50	Instance recall @ IoU 0.50	Count MAE
BubbleID base	0.6772	0.8076	0.6354	0.3895	33.86
BubbleID-Flow	0.0585	0.1105	0.6592	0.1037	72.13

This is an off-the-shelf cross-domain screening result, not a claim of generalized validation. Each checkpoint was applied outside the domain it was fitted for, so these numbers characterize transfer behavior under the stated protocol and should not be read as a ranking of method quality.

Out-of-scope use

Do not use this version to claim validated three-dimensional vapor volume, void fraction, tracking accuracy, bubble lifetime, departure frequency, or coalescence or breakup event accuracy. Those require calibration and temporal identity or event annotations that are not part of this version.

Use rules

1. Preserve raw/ unchanged.
2. Use only the supplied grouped split files for reported benchmark results.
3. Do not split frames from the same (regime, source video/power condition) across partitions.
4. Report performance by regime, not only a pooled score.
5. Treat annotations flagged in metadata/annotation_qa.md according to the benchmark evaluator's documented ignore policy.

Attribution

Cite the BoilingBench-CV dataset version, the Zenodo DOI 10.5281/zenodo.22264378, and the source records listed in metadata/data_sources.csv.

Upstream sources

The deposit records these relations. Honor them alongside the CC BY 4.0 attribution requirement:

Relation	Resource
Is derived from	Dryad 10.5061/dryad.ksn02v7gx
Is derived from	Dryad 10.5061/dryad.kh18932mw — also mirrored at UARK-NED3/PoolBoiling-HighSpeedVideo
Is supplement to	github.com/UARK-NED3/BoilingBench-CV
References	github.com/cldunlap73/BubbleID

Overlap with PoolBoiling-HighSpeedVideo

The 10.5061/dryad.kh18932mw deposit is mirrored on the Hub as UARK-NED3/PoolBoiling-HighSpeedVideo, and the two repositories share content: all 357 pool-boiling annotated images, all four BubbleContours.json files, and 8 of the PSi-HFE videos here also appear there.

Use this repository for the canonical COCO-style annotations, the grouped benchmark splits and their leakage rule, and the segmentation evaluation records. Use PoolBoiling-HighSpeedVideo for the complete high-speed video set (31 videos across four regimes), the original contour JSON format, and the MATLAB descriptor toolkit. The licenses differ — this package is CC BY 4.0, that deposit is CC0 1.0 — so check the terms for the copy you actually use.

Source method papers:

• C. Dunlap et al., "BubbleID: A Deep Learning Framework for Bubble Interface Dynamics Analysis," Journal of Applied Physics 136, 014902 (2024).
• A. Fahim et al., "BubbleID-Flow: Machine-Vision Quantification of Vapor Area Fraction in Subcooled Flow Boiling," manuscript in preparation.

The CC BY 4.0 license applies to this package as organized here. It does not erase attribution obligations or source-specific terms recorded in metadata/data_sources.csv and metadata/data_rights_manifest.csv.